Model card: drop the results section entirely
Browse files
README.md
CHANGED
|
@@ -64,23 +64,6 @@ Each server publishes its deployment contract in the connect-time metadata
|
|
| 64 |
> Piper is 30 Hz. Actions are absolute (G2) or delta-from-start (Piper) joint targets;
|
| 65 |
> stepping G2's chunk at 30 Hz would execute it 3× too fast.
|
| 66 |
|
| 67 |
-
## Results (G2)
|
| 68 |
-
|
| 69 |
-
**G2** — offline replay against a live server, 120 frames / 12 episodes:
|
| 70 |
-
|
| 71 |
-
| metric | value | threshold |
|
| 72 |
-
|---|---|---|
|
| 73 |
-
| joint MAE | **0.28°** (worst 0.95°) | < 5° |
|
| 74 |
-
| gripper agreement | **99.8%** | > 90% |
|
| 75 |
-
| delta direction cosine | **0.999** | > 0.8 |
|
| 76 |
-
|
| 77 |
-
Validated on hardware in four stages — dry-run → ground-truth replay → policy-replay on
|
| 78 |
-
dataset observations → live closed loop — and completes the task in the last stage.
|
| 79 |
-
|
| 80 |
-
**Piper** is included as a **worked example**, not as a results claim: a second robot with a
|
| 81 |
-
different arm count, action space and control rate, so the pipeline can be compared against
|
| 82 |
-
something other than G2. No success-rate numbers are reported for it here.
|
| 83 |
-
|
| 84 |
## Things that will bite you
|
| 85 |
|
| 86 |
* **G2 gripper polarity is inverted between dataset and wire.** In the dataset `1.0` = OPEN,
|
|
|
|
| 64 |
> Piper is 30 Hz. Actions are absolute (G2) or delta-from-start (Piper) joint targets;
|
| 65 |
> stepping G2's chunk at 30 Hz would execute it 3× too fast.
|
| 66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
## Things that will bite you
|
| 68 |
|
| 69 |
* **G2 gripper polarity is inverted between dataset and wire.** In the dataset `1.0` = OPEN,
|